基于细粒度信任推导的社交推荐模型研究

发布时间：2018-01-25 21:51

本文关键词： 推荐系统协同过滤细粒度信任推导最近邻算法矩阵分解　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随信息数量的高速增长,信息过载成为了当前互联网环境下无法避免的问题,快速、准确地获取有效信息变得越来越困难。推荐系统能够有效地挖掘用户特征、捕获用户兴趣,并在对用户偏好进行建模的基础上主动为用户提供个性化的信息过滤和商品推荐等服务。无论是在信息检索还是电子商务领域,推荐系统已经成为应对信息过载,提升用户满意度的有效方法。协同过滤凭借其计算简便以及推荐准确率高等优势,成为了应用最为广泛的推荐技术之一。然而,其推荐性能在很大程度上受到了用户反馈的稀疏性和冷启动等问题的影响。社交网络的出现和发展为推荐系统研究注入了新的活力,用户不再被视作独立的个体,他们通过社交关系彼此相连,相互影响。信任是社交关系的核心与关键,构建基于信任的社交推荐系统,不仅能缓解用户反馈稀疏、提高推荐的准确率和覆盖率,也为用户提供了更加可信的推荐结果,提高了系统安全性。本文围绕如何进一步提升基于信任的社交推荐系统的性能,提出了一种综合多种信任属性的细粒度信任推导模型,并据此对协同过滤算法的两类经典实现进行扩展,构建了基于细粒度信任推导的社交推荐模型。结合信任的传递性、相似性和多维性,我们构建了一个细粒度信任推导模型。首先,提出了一种基于关键用户挖掘的单步信任传播方法,并引入相似度模型进行推导优化,扩展了信任网络覆盖率;接着,通过用户间交互频率和反馈进行信任粒度度量;最后,引入多维信任推导方法,利用影响信任关系的隐式因子对缺失的细粒度信任进行进一步推导。通过实验,验证了该信任推导模型能够有效地解决社交关系稀疏及社交粒度单一等问题。以该信任推导模型为基础,本文对最近邻算法和隐语义模型进行扩展,提出了基于细粒度信任推导的近邻模型NNST和社交联合矩阵分解模型CMFST。通过多组对比实验,验证了本文提出的社交推荐模型能够有效地解决传统个性化推荐系统中存在的数据稀疏性和冷启动等问题,在提高了推荐系统的覆盖率的同时也为用户提供了更加准确的社交推荐结果。
[Abstract]:With the rapid growth of the amount of information, information overload has become an unavoidable problem in the current Internet environment. It is becoming more and more difficult to obtain effective information accurately. Recommendation system can effectively mine user features and capture user interest. And on the basis of user preference modeling, it provides users with personalized information filtering and commodity recommendation services, whether in the field of information retrieval or e-commerce. Recommendation system has become an effective method to deal with information overload and improve user satisfaction. Collaborative filtering has the advantages of simple calculation and high recommendation accuracy. Has become one of the most widely used recommended technologies. The performance of recommendation is greatly affected by the problems of user feedback sparsity and cold startup. The emergence and development of social networks inject new vitality into the research of recommendation system. Users are no longer regarded as independent individuals, they are connected and influence each other through social relationships. Trust is the core and key of social relations. It can not only alleviate the sparse feedback of users, improve the accuracy and coverage of recommendations, but also provide users with more reliable recommendation results. This paper proposes a fine-grained trust derivation model based on multiple trust attributes, focusing on how to further improve the performance of trust based social recommendation system. Based on this, two classical implementations of collaborative filtering algorithm are extended, and a social recommendation model based on fine-grained trust derivation is constructed, which combines the transitivity, similarity and multi-dimension of trust. We construct a fine-grained trust derivation model. Firstly, a one-step trust propagation method based on key user mining is proposed, and the similarity model is introduced to derive and optimize the trust network coverage. Then, the granularity of trust is measured by the interaction frequency and feedback between users. Finally, the multi-dimensional trust derivation method is introduced to further deduce the missing fine-grained trust by using the implicit factors which affect the trust relationship. It is verified that the trust derivation model can effectively solve the problems of sparse social relationship and single social granularity. Based on the trust derivation model, this paper extends the nearest neighbor algorithm and hidden semantic model. The nearest neighbor model (NNST) based on fine-grained trust derivation and the social joint matrix decomposition model (CMFST) are proposed. It is verified that the proposed social recommendation model can effectively solve the problems of data sparsity and cold start in the traditional personalized recommendation system. It not only improves the coverage of recommendation system, but also provides users with more accurate social recommendation results.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】